fix(v3): prompt perturbation to break deterministic outputs

fix(v3): include multi-step tasks from curriculum start

feat(v3): custom multi-step GRPO training loop with GiGPO

feat(v3): add rollout logic and fix target_crease_edges bug

Add MAX_CONCURRENT_ENVS, sync latest changes

Rename server/ to origami_server/ to avoid module name conflict with uvicorn.server

Origami RL environment with Three.js viewer